Using Machine Learning for Early Recognition and Personalized Treatment of Acute Kidney Injury

点击翻译按钮获取中文摘要

PROJECT SUMMARY Acute kidney injury (AKI) occurs in up to 20% of hospitalized patients and is associated with increased risk of readmission, morbidity, and mortality. The estimated annual cost of AKI care in the US is over 10 billion dollars, and, with the incidence rising, these costs will continue to increase. The current gold standards for diagnosing AKI, creatinine and urine output, are often delayed in their recognition of tubular injury. Prior work on AKI has typically focused on patients who have already developed AKI based on these standards, and interventions at this late time point have had mixed success. In contrast, emerging data suggest that intervening earlier can improve outcomes. Therefore, it is critical to optimize the early detection of AKI in hospitalized patients. We have previously developed a machine learning tool to identify patients at high risk of severe (stage 2 or greater) AKI more than a day earlier than clinically apparent using structured electronic health record (EHR) data. Although more accurate than prior methods, it suffers from a high rate of false positives, which limits its value in clinical practice. There is a large amount of valuable information that is stored in unstructured free-text fields (e.g., clinical notes) that could be utilized using natural language processing (NLP) within advanced deep learning neural network models that could significantly improve the detection of early AKI. Furthermore, there are established and emerging kidney injury biomarkers that could be combined with EHR-based models to improve accuracy even further. Finally, it remains unclear what interventions will have the best chance of decreasing the risk for developing severe AKI in high-risk patients. A better understanding of which interventions are of greatest benefit to specific patients is critical for improving the outcomes of patients at risk of AKI. The objective of this project is to develop novel tools to improve the identification and treatment of patients at high risk of AKI using a large, multicenter cohort. In Aim 1, we will use NLP and deep learning algorithms to develop a model to predict severe AKI across four health systems. In Aim 2, we will silently run the best- performing model developed in Aim 1 in real-time to identify high-risk patients. Manual retrospective chart review will be performed on a cohort of the highest risk patients to determine both the proportion of patients who receive guideline-based care as well as the association between receipt of guideline-based care and outcomes. We will also identify novel phenotypes of patients who are particularly helped or harmed by specific guideline-based interventions. Finally, in Aim 3, we will collect kidney injury biomarkers in the highest-risk patients to determine the added value of biomarkers to EHR-based models alone. Our proposal will provide clinicians with new tools to identify patients at risk of AKI earlier and more accurately. It will also provide evidence for which interventions are most likely to improve patient outcomes. This will result in earlier, more personalized care for patients at high risk of AKI, which will lead to decreased costs, morbidity, and mortality.